The FastDownward Benchmark

Task: select heuristics for the FastDownward planner
Cost: number of optimization steps
Number of hyperparameters to control: one categorical
State Information: average value, max value, min value, number of open list entries and variance for each heuristic
Noise Level: fairly large
Instance space: either specifically desigd easy toy instances with ground truth or common planning competition instance sets

This benchmark is an interface to the Fast Downward AI planner, controlling its heuristic hyperparameter. In each step until the algorithm finishes or is terminated via the cutoff, the DAC controller selects one of either two (toy case) or four heuristiccs for the planner to use. The goal is to reduce the runtime of the planner, so every step that is taken in the benchmark incurs a cost of 1.

Out of our real-world benchmarks, FastDownward is likely the fastest running and it has been shown to be suitable to dynamic configuration. Though the noise level is fairly high, most DAC controllers should be able to learn functional policies in a comparatively short time frame.

The FastDownward benchmark was constructed by Speck et al. for the paper `”Learning Heuristic Selection with Dynamic Algorithm Configuration” <https://arxiv.org/pdf/2006.08246.pdf>`_ at ICAPS 2021

class dacbench.benchmarks.fast_downward_benchmark.FastDownwardBenchmark(config_path=None, config=None)

Bases: AbstractBenchmark

Benchmark with default configuration & relevant functions for Sigmoid

get_benchmark(seed=0)

Get published benchmark

Parameters: seed (int) – Environment seed
Returns: env – FD environment
Return type: FastDownwardEnv

get_environment()

Return Luby env with current configuration

Returns: Luby environment
Return type: LubyEnv

read_instance_set(test=False): Read paths of instances from config into list

Planning environment from “Learning Heuristic Selection with Dynamic Algorithm Configuration” by David Speck, André Biedenkapp, Frank Hutter, Robert Mattmüller und Marius Lindauer. Original environment authors: David Speck, André Biedenkapp

class dacbench.envs.fast_downward.FastDownwardEnv(config)

Bases: AbstractEnv

Environment to control Solver Heuristics of FastDownward

close()

Close Env

Returns: Closing confirmation
Return type: bool

kill_connection(): Kill the connection

recv_msg()

Recieve a whole message. The message has to be prepended with its total size Based on comment from SO see [1]

Returns: The message as byte
Return type: bytes

recvall(n: int)

Given we know the size we want to recieve, we can recieve that amount of bytes. Based on comment from SO see [1]

Parameters: n (int) – Number of bytes to expect in the data
Returns: The message as byte
Return type: bytes

render(mode: str = 'human') → None

Required by gym.Env but not implemented

Parameters: mode (str) – Rendering mode

reset()

Reset environment

Returns: State after reset
Return type: np.array

send_msg(msg: bytes)

Send message and prepend the message size

Based on comment from SO see [1] [1] https://stackoverflow.com/a/17668009

Parameters: msg (bytes) – The message as byte

step(action: Union[int, List[int]])

Environment step

Parameters: action (Union[int, List[int]]) – Parameter(s) to apply
Returns: state, reward, done, info
Return type: np.array, float, bool, dict

class dacbench.envs.fast_downward.StateType(value)

Bases: Enum

Class to define numbers for state types